Turk J Elec Eng & Comp Sci
(2018) 26: 974 – 986
c⃝ TUBITAK
doi:10.3906/elk-1611-172
Turkish Journal of Electrical Engineering & Computer Sciences
http :// journa l s . tub i tak .gov . t r/e lektr ik/
Research Article
Transient- and probabilistic neural network-based fault classification in EHV
three-terminal lines
Ravi Kumar Varma BHUPATIRAJU∗, Venkata Sesha Samba Siva Sarma DHANIKONDA,Venkata Ramana Rao PULIPAKA
Department of Electrical Engineering, National Institute of Technology, Warangal, India
Received: 17.11.2016 • Accepted/Published Online: 01.08.2017 • Final Version: 30.03.2018
Abstract: This paper presents a fast and accurate fault classifier for three-terminal transmission circuits. Traditional
phasor-based methods fail to meet the high speed requirements of modern power grids and necessitate alternative
solutions. The transient-based schemes use advanced signal processing techniques to achieve fast and accurate fault
classification. As the three-terminal lines experience very pronounced transients during faults, the proposed method
makes use of the fault-generated transients to quickly and correctly classify the fault. Many transient-based schemes
fail to give the required accuracy since the transient patterns with relay-measured signals are highly influenced by fault
conditions. Therefore, a thorough analysis of transient patterns is carried out in this paper, and based on the typical
patterns revealed by the analysis of fault-generated transients an effective classification algorithm is developed. For
high-speed classification, only a quarter-cycle of postfault voltage signals measured at the relay points will be processed
for feature extraction using wavelet transform. The algorithm includes a hybrid procedure based on a probabilistic neural
network for tackling the effects of fault inception angle and fault resistance in transient variations. Particularly, it is
designed to overcome the problem with double-line-to-ground fault classification. The technique is simple and extensive
simulation studies and comparison substantiate the efficacy of the proposed method under different fault conditions.
Key words: Three-terminal lines, fault classification, transients, wavelet transform, probabilistic neural network
1. Introduction
Modern power grids are becoming complex interconnected power systems to facilitate the optimal utilization
of existing infrastructure and resources. In such power systems, extra-high voltage transmission lines carry
the bulk of the power to meet the demand at different load centers. Transmission lines play a vital role in
maintaining the grid integrity and therefore must have very effective protective relays for the reliable and stable
operation of the grid. Generally, a three-terminal line configuration is created if construction of a new substation
is not possible due to cost or right-of-way concerns. Significantly, due to lack of measurements at the tap point,
a three-terminal line protection requires further enhancements [1]. Figure 1 shows such a 400 KV line structure
present in an Indian power grid [2].
In transmission line protection, after fault detection, the line relay invokes a fault classification function
to identify the type of fault and the faulted phases for initiating possible single-pole tripping and autoreclosing
[3]. Hence, the fault classifier must be very quick and accurate to maintain system stability and continuity of
supply. Many fault classification algorithms are traditionally phasor-based, where the estimation of postfault
∗Correspondence: [email protected]
974
BHUPATIRAJU et al./Turk J Elec Eng & Comp Sci
1T
3
2
S130 KM
70 KM
60 KMS2
S3
Section 1
Section 3
Section 2
Figure 1. A 400 KV three-terminal transmission circuit.
voltage/current phasors requires at least a half-cycle of data. Most of these methods are affected by dynamically
changing network configuration and fault conditions. Recently, a phasor-based method using sequential reactive
powers was proposed by Mahamedi and Zhu [4]. Despite its claim to be setting- and synchronization-free, one
entire cycle of fault data requirement is a critical factor influencing the speed of response.
Driven by the need for high-speed relaying, with the availability of efficient signal-processing techniques
and high-sampling-rate hardware a new class of classification algorithms based on fault-induced transients has
evolved. These transient-based algorithms process fault-generated traveling waves or high-frequency components
for feature extraction, mostly by using wavelet transform. In [5], a method using an initial current traveling
wave is reported, mentioning that it fails during three certain phase faults. A wavelet singular entropy-based
technique is presented in [6] by using a half-cycle postfault voltage signal. The problem with this method
is it requires the setting of three thresholds according to system situation. Another current transient-based
classification technique is proposed in [7], whose accuracy with double-line to ground (LLG) faults is only around
90%. Despite their speed advantage, a major problem for all these deterministic transient-based methods is the
influence of fault inception angle (FIA) and fault resistance (Rf ) on the fault transient magnitudes. Particularly,
LLG fault classification poses a big challenge due to misleading transient patterns generated during the faults.
In an effort to overcome the above problems due to system inconstancy and for better performance,
several authors explored artificial intelligence applications with fuzzy logic [8,9], artificial neural networks
(ANNs) [10,11], support vector machines (SVMs) [12,13], etc., to solve the fault classification problem. The
inputs to these knowledge-based methods are mainly either steady-state or transient components of the fault
voltage/current signals. Recently, Jamehbozorg and Shahrtash proposed a decision tree (DT)-based method
with half-cycle discrete Fourier transformation (HCDFT) [14], which requires computation of ten odd harmonic
phasors. A statistical DT-based approach [15] is accurate, but it takes two full cycles of current signals for
transient extraction. Another transient-ANN approach was developed with a rough membership neural network
in [16], but the results presented include only a 90 variation of FIA. In addition to all the above methods
primarily developed for two-terminal lines, a transient SVM-based classifier is designed for three terminal lines
in [17]. However, its poor performance (95% accuracy) in classifying grounded faults is not acceptable. In view
of the above drawbacks, a thorough investigation into the transient patterns generated during faults is necessary
for developing a useful classifier.
In this paper, a new transient-based scheme for fast and accurate fault classification is proposed for
three-terminal lines. The proposed algorithm extracts the distinguishing features from the three phase voltage
signals by using wavelet transformation and a probabilistic neural network (PNN) is included to improve the
accuracy of the classifier with LLG faults. Unlike many other techniques, the scheme is simple, requires single
threshold setting, and its speed and accuracy are significantly high. The performance of the method is verified
by simulation of a three terminal line shown in Figure 1 using MATLAB/SIMULINK environment.
975
BHUPATIRAJU et al./Turk J Elec Eng & Comp Sci
In Section 2, the feature extraction is described, followed by feature analysis and algorithm formulation.
Section 3 presents the simulation details and performance results of the proposed algorithm. Finally, conclusions
are given in Section 4.
2. Transient features and algorithm formulation
Fault classification methods process voltage signals and/or current signals measured by the relays to extract
necessary discriminating features. Since most of the available algorithms use only postfault signals, the size of
a signal window plays an important role in deciding the speed and accuracy of the classification procedure. A
larger window size normally improves accuracy at the cost of speed and vice versa. However, use of a smaller
window without losing much accuracy is possible in transient-based schemes because of the wide-band nature
of the fault transients. In the proposed method, only postfault voltage signals of a quarter cycle duration from
all the three ends are taken as inputs and the choice is based on the fact that voltage transients are more than
current transients during faults in power systems [18]. A sampling frequency of 200 kHz is used as the faulted
lines contain high frequency components in the range of 10–100 kHz [19]. Such a sampling rate is common with
currently available hardware and, in addition, avoids the use of an antialiasing filter in digital relays [20]. The
high-frequency components present in voltage signals are extracted by using discrete wavelet transform (DWT),
which has proven to be an effective signal processing tool for power systems [21].
The feature extraction procedure is presented below, followed by a discussion on fault transient patterns
for developing the classification procedure.
2.1. Feature extraction
Figure 2 depicts the overall structure of the proposed scheme in which the classifier is activated by the fault
detector when a fault is detected in the line. The classifier algorithm takes the three phase voltage signals
(Va , Vb , and Vc) of 1/4 cycle time from the three ends and forms the line voltages (Vab , Vbc , Vca) and
zero-sequence voltage (V0) from the phase voltages. In spite of the communication requirements, use of three
end signals improves the feature strength, as explained later. The quarter-cycle windows of all fault voltage
signals shown in Table 1 are then decomposed to the first level by using the multiresolution analysis feature of
the DWT.
T
Communication Lines
S 1
S 3
S 2
Fault Detector
Fault Classifier
20 GVA 10 GVA
5 GVA
PT
PT
PT
Figure 2. Signal communication for the proposed fault classifier.
976
BHUPATIRAJU et al./Turk J Elec Eng & Comp Sci
Table 1. Voltages decomposed by the DWT.
End Phase voltages Line voltages Zero sequence voltages1 Va1, Vb1, Vc1 Vab1,Vbc1,Vca1 V01 = (Va1 + Vb1 + Vc1)/32 Va2, Vb2, Vc2 Vab2,Vbc2,Vca2 V02 = (Va2 + Vb2 + Vc2)/33 Va3, Vb3, Vc3 Vab3,Vbc3,Vca3 V03 = (Va3 + Vb3 + Vc3)/3
The DWT of a digital signal is given by
DWT (m, k) =1
√am0
∑n
x(n)g
(k − nb0a
m0
am0
), (1)
where g(n) is the mother wavelet and x(n) is the input signal, and the scaling and translation parameters a
and b are functions of integer parameter m . Figure 3 shows measured phase voltage signals at end-1 for an
AG fault with the fault inception at 0.025 s. A 5 ms postfault voltage signal window and the first-level detail
coefficients (D1) coefficients of phase and zero-sequence voltages obtained using wavelet transformation are also
indicated.
Although many earlier techniques used higher-level decomposition, in this paper only first-level is pre-
ferred due to computational simplicity and accuracy of transient energy calculation. The transient-based meth-
0 0.01 0.02 0.03 0.04 0.05 0.06-5
0
5x 105
Vo
l ts
0.025 0.0255 0.026 0.0265 0.027 0.0275 0.028 0.0285 0.029 0.0295 0.03-5
0
5x 105
Vo
lts
0 .025 0.0255 0.026 0.0265 0.027 0.0275 0.028 0.0285 0.029 0.0295 0.03-5
0
5x 104
D1
0 .025 0.0255 0.026 0.0265 0.027 0.0275 0.028 0.0285 0.029 0.0295 0.03-5
0
5x 104
D1
0 .025 0.0255 0.026 0.0265 0.027 0.0275 0.028 0.0285 0.029 0.0295 0.03-5
0
5x 104
D1
0 .025 0.0255 0.026 0.0265 0.027 0.0275 0.028 0.0285 0.029 0.0295 0.03-5
0
5x 104
Time
D1
V0 1
Va1 Vb 1 Vc 1
Quarter Cycle Pos t -fault data Window
D1 coefficients of Zero sequence voltage at end 1
Figure 3. Phase voltages at end 1 and the D1 coefficients with an AG fault in Section 3.
977
BHUPATIRAJU et al./Turk J Elec Eng & Comp Sci
ods need denoising of input signals before their processing to avoid the effect of noise and an assumption is
made to that effect with the proposed method in this paper.
From the D1 coefficients, the transient signal energy for each voltage signal is calculated by using the
equation
Transient energy ψ =
√√√√ N∑i=1
[D1 (i)]2, (2)
where N is the number of first-level detail coefficients after truncation of five samples, applied to eliminate the
edge effect.
Now the effective transient energy for each voltage signal is obtained by adding the signal energies of
the three ends (1, 2, and 3). The effective transient energy for phase, line, and zero-sequence voltage signals is
given by
Ψi = Ψi1 +Ψi2 +Ψi3 ; i = abc
Ψjk = Ψjk1 +Ψjk2 +Ψjk3 ; j = abc& k = bca
Ψ0 = Ψ01 +Ψ02 +Ψ03,
where Ψi is the effective transient energies of phase voltages, Ψjk is the effective transient energies of line
voltages, and Ψ0 is the effective transient energy of the zero-sequence voltage. The above set of energies is
normalized to form a feature array, called the normalized effective transient energy (NETE) array, as shown in
Figure 4. The NETE array, describing the high-frequency transient content on voltage signals, is used as input
for classifying faults in the present technique.
Zero Sequence Voltage
Ψ0ΨbΨa Ψc Ψab Ψbc Ψca
Phase Voltages Line Voltages
Figure 4. NETE array input to the classifier.
In the above procedure, the addition of three end energies is done to improve the sensitivity of the scheme
with respect to the fault location and Rf . In close-up faults, the voltage signal of the faulted phase measured
by the relay (at the faulted section end) contains very little transient energy. However, the relays at the other
two ends find appreciable energy in the measured signals. Therefore, use of only single-end measurements leads
to errors in classification of close-up faults.
2.2. Analysis of transient patterns:
Many transient-based classification techniques depend on a comparison of the fault transient energy magnitudes
of different phase voltages or currents and these magnitudes widely vary depending on the type of fault and fault
parameters. Therefore, methods based on the direct energy threshold fail to give accuracy and normalization
becomes a necessity to fix the thresholds. Despite the normalization, threshold fixing is difficult and often leads
978
BHUPATIRAJU et al./Turk J Elec Eng & Comp Sci
to inaccurate classification. An advantage of the proposed method in this paper is that it requires a single
threshold fixing.
The NETE feature array formed from the fault voltage signals will provide the necessary information for
classifying the fault. Initially, the ground involvement in the fault can be checked by verifying the energy on
zero sequence voltages, i.e. Ψ0 . For ungrounded faults, its value will be zero.
2.2.1. Grounded faults
Single line-to-ground (LG) faults: In LG faults, which amount to 85% of total faults in power systems, the
unfaulted phases also contain energy magnitudes that are close to that of the faulted phase. The unfaulted
phase energies (Ψb and Ψc) are nearer to the faulted phase energy (Ψa), thus making segregation of the faulted
phase difficult. Instead of using phase voltage energies, the proposed method uses energy on line voltages to
resolve this problem. Observing the energy on line voltages, Ψbc becomes zero, clearly indicating that the fault
is on phase A. Similarly, Ψca and Ψab will become zero for BG and CG faults, respectively.
LLG faults: Accuracy of classification in many algorithms is mainly affected by LLG faults, which
create misleading energy patterns. Although in many situations the faulted phases carry higher magnitudes of
transients than the unfaulted phases, during certain ranges of FIA, depending on the Rf , the transient presence
occurs the other way. Further, Figure 5 demonstrates the variation in energy with respect to FIA on phase
voltages during a BCG fault. The energy on faulted phases is more than the unfaulted for many inception
angles, except in the band from 60 to 130 . The width of this band is again a function of the Rf . Similar
inconsistent patterns appear for ABG and CAG faults, also with error bands between 0 and 180 . Nearly 1/3
of the LLG faults will be misclassified if the decision is based on the comparison of energy present on individual
phases. Hence, a neural network is considered a preferable solution for LLG fault classification in this work.
0 10 30 5000
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Ψa Ψb Ψc
NE
TE
of
Pha
s e V
olta
ges
Fault Inception Angle
BCG Fault in S ection 3 L=60KM Rf =100 Ohms
70 90 110 130 150 170
Figure 5. Variation of energy on phase voltages with FIA (BCG fault).
2.2.2. Ungrounded faults
In the case of line-to-line (LL) faults, the two faulted phases contain equal values of transient energy in their
phase voltages and the other phase contains no transient energy. This distinctive feature clearly segregates an
LL fault and the ground is not involved in the fault. If the above condition is not satisfied during a fault without
ground involvement, then the fault is decided to be a three-phase (LLL) fault.
979
BHUPATIRAJU et al./Turk J Elec Eng & Comp Sci
2.3. Algorithm formulation
Based on the above feature analysis, a hybrid algorithm is developed in this paper to classify all types of faults
very accurately. In this algorithm, determination of ground involvement and classification of LG, LL, and LLL
faults will be carried out by a deterministic procedure, as explained earlier. For LLG faults, in view of the
misleading transient patterns, a neural network is incorporated, thus making the algorithm a hybrid one. The
algorithm takes the NETE array as input and initially checks whether the fault is grounded or not. If the fault
does not involve the ground, then it checks for LL or LLL faults. In the case of a grounded fault, the algorithm
first checks for different LG faults; if not concluded, the only remaining possibility is an LLG fault. Then a
trained PNN is activated for LLG faults to identify the faulted phases. A PNN application for only LLG faults
gives better overall accuracy and makes the algorithm suitable for real-time implementation. A brief description
of PNNs followed by the overall flowchart is given below.
PNNs: a PNN is a supervised learning network that belongs to the general category of radial basis
networks. Many classification problems utilize pattern recognition techniques that are based on the theory of
Bayesian classification [22]. These techniques, realized by probabilistic model networks, require estimation of
probability density function (PDF) and for an input x belonging to a category A. The PDF is given by
fA (X) =1
(2π)p/2σp
1
m
m∑1
exp
[− (X −XAi)
T(X −XAi)
2σ2
](3)
where i is the pattern number, m is the total number of training patterns, XAi is ith pattern from category A,
σ is the smoothing parameter, and p is the dimensionality of measurement space. The smoothing parameter is
fixed by the spread value in MATLAB simulation.
With enough training data, a PNN is guaranteed to converge and its fast learning makes it suitable for
real-time applications, as in [23]. In the present application, the PNN is initially trained with several NETE
patterns obtained from the simulation for different LLG faults. Patterns are generated by considering wide
variations in FIA, Rf , and fault location. The PNN will have seven inputs and a single output, as shown
in Figure 6a. The overall flowchart for the proposed scheme is shown in Figure 6b. The value of ε (single
threshold) is set to be very small and should be ideally zero.
3. Simulation details and results
The classification procedure developed above was verified by simulation using the 400 KV line shown in Figure
1 with the parameters given in Table 2. The system was simulated in MATLAB/SIMULINK with a distributed
parameter model (Bergeron’s model) for the transmission lines, so that the fault signals exhibited the high-
frequency transient components generated. Wavelet decomposition was done using bior2.2 as the mother wavelet
because of its effectiveness in processing power system transients [24]. A total of 14,040 different fault cases
were considered for simulation by varying fault type, fault location, FIA, and Rf , as shown in Table 3, to verify
the performance of the proposed algorithm with a practical value of 0.02 for ε . The transient energy patterns
obtained for different faults are shown in polar scatter plots to justify the validity of the proposed algorithm.
Table 2. 400 KV Transmission line parameters.
R1 = 0.029792 Ω/km L1 = 1.05678 mH/km C1 = 11.04137 nF/kmR0 = 0.16192 Ω/km L0 = 3.947042 mH/km C0 = 7.130141 nF/km
980
BHUPATIRAJU et al./Turk J Elec Eng & Comp Sci
Ψa
Ψb
Ψa b
Ψ c
Ψ c a
Ψ0
Proba bilis tic Ne ura l
Ne tw
(a)
(b)
orkΨbc
Input NETE Arra y
Output4: AB G5: B CG6: CAG
Y
Y
Is Ψbc < ε ?
START AFTER FAULT DETECTION
Take quarter cycle data windows of all voltages from three ends
Ca lc ula te e ne rg ie s for a ll volta g e s
1AG
Fa ult
7AB
Fa ult
PNN for LLG Fa ult
Cla s s ific a tion
YY
Y
Y
Y
Sum Energies of three ends and obtain NETE
( )[ ]Σ=
-N
i
iD1
21ψ
[Ψa Ψb Ψc Ψab Ψbc Ψca Ψ0]
Is Ψ c a < ε ?
2B G
Fa ult
Is Ψa b < ε ?
3CG
Fa ult
Is Ψ c < ε & Ψa=Ψb?
8B C
Fa ult
Is Ψa < ε & Ψb=Ψ c?
9CA
Fa ult
Is Ψb < ε & Ψa=Ψ c?
Is 3Ψ0< ε ?
10AB C Fa ult
N
4AB G Fa ult
5B CG Fa ult
6CAG Fa ult
Grounde d fa ultUng rounde d fa ult
N N
Apply DWT to find Detail coefficients of
Va, Vb, Vc ,Vab, Vbc , Vca a nd V0
Figure 6. a) Inputs and outputs for the PNN. b) Overall flowchart for the proposed algorithm.
981
BHUPATIRAJU et al./Turk J Elec Eng & Comp Sci
Table 3. Fault conditions considered.
Condition RangeFault type AG, BG, CG, AB, BC, CA, ABG, BCG, CAG, ABCFault resistance 0 Ω, 50 Ω, 100 ΩFault inception angle 0 to 350 in steps of 10
Fault location Every 10 KM in each sectionTotal number of cases 14,040 (1404 for each fault type)
The performance of the proposed technique for LG faults was assessed by simulating 4212 faults at
different locations. Figures 7a and 7b show the variation of transient energy with FIA on phase and line
voltages, respectively, for AG faults (1404 cases). Note that Ψb and Ψab patterns are not visible in these plots
as they lie beneath Ψc and Ψca , respectively. Although the energy values of phase voltages for both faulted
(Ψa) and unfaulted (Ψb and Ψc) lines are nearly overlapping, the values of line voltages clearly indicate the
faulted phase; i.e. for all AG faults, the value of Ψbc approaches zero. The present procedure had an accuracy
rate of 100% in classifying LG faults.
0.2
0.4
0.6
30
210
60
240
90
270
120
300
150
330
180 0
o -- Ψa o – Ψb o - Ψc
FIA 0.2
0.4
0.6
30
210
60
240
90
270
120
300
150
330
180 0
o -- Ψa b o – Ψbc o - Ψc a
FIA
Figure 7. Energy patterns for AG faults a) phase b) line.
In LLG faults, the transients generated clearly brought out the typical patterns that make classification
difficult. Figures 8a and 8b show the variation of energy on phase voltages for ABG faults with two different
Rf s (0 and 100 Ω). The unfaulted phase energy (Ψc) was more than the faulted phase energy for a range of
inception angles and this range was large, with an increase in Rf . Similar patterns were obtained for BCG
and CAG faults as well. The symmetry of patterns in the polar plots shows that consideration of FIA variation
from 0 to 180 is very much necessary while developing transient-based classification algorithms.
Regarding performance levels, the PNN was activated for all LLG faults without fail, and the classification
accuracy with the trained PNN was very high (99.59%). The results of the 4212 cases simulated validate the
application of PNN for classifying only LLG faults.
In the LL faults, the present technique classified all faults except 2 (out of 4212 cases) correctly by using
the energy values of the phase voltages. Figure 9a shows the variation in energy of phase voltages for all phase
A-to-phase B (AB) faults simulated (Ψa patterns not visible). Clearly, the unfaulted phase always exhibited
982
BHUPATIRAJU et al./Turk J Elec Eng & Comp Sci
0.2
0.4
0.6
30
210
60
240
90
270
120
300
150
330
180 0
o -- Ψa b o – Ψbc o - Ψc a
FIAFIA
180
0.2
0.4
0.6
30
210
60
240
90
270
120
300
150
330
0
o -- Ψa o – Ψb o - Ψc
Figure 8. Energy patterns on phase voltages for ABG faults a) Rf = 0 Ω, b) Rf = 100 Ω.
a value of energy (Ψc) close to zero. The normalized energy on faulted phase voltage signals (Ψa and Ψb)
remained fairly constant at 0.353. A similar pattern was obtained for BC and CA faults as well.
o -- Ψa o – Ψb o -- Ψc
0.2
0.4
0.6
30
210
60
240
90
270
120
300
150
330
180 0
FIA
90
180
0.2
0.4
0.6
30
210
60
240
270
120
300
150
330
0
FIA
o -- Ψa o – Ψb o -- Ψc
Figure 9. Energy patterns on phase voltages a) AB faults, b) ABC faults.
All the symmetrical three-phase faults (ABC faults) simulated were classified by the present method
with 100% accuracy. Figure 9b shows the typical transient patterns developed on phase voltages for three-phase
faults. The NETE magnitudes of different voltage signals varied widely with respect to FIA and in no case did
the NETE meet the pattern of any LL fault.
The overall accuracy of performance with the proposed hybrid classifier was very high. Particularly,
its ability to classify more frequent LG faults with 100% accuracy is an attractive feature. The results of
the simulation show that the technique is quite insensitive to fault location, Rf , and FIA, which is a definite
advantage over conventional phasor-based methods. Moreover, the classification of LLG faults using a PNN
overcomes the major drawback of many existing transient-based methods in classifying LLG faults. Table 4
gives the complete simulation results for all 14,040 faults considered, with only 17 misclassifications in total. A
983
BHUPATIRAJU et al./Turk J Elec Eng & Comp Sci
performance comparison with three recently reported methods [7,17,25] in Table 5 reveals the improvement in
accuracy achieved by the proposed technique.
Table 4. Classification results.
Given fault type Fault classification result
(All 1404 cases)↓ 1 AG 2 BG 3 CG 4 ABG 5 BCG 6 CAG 7 AB 8 BC 9 CA 10 ABC
AG 1404 0 0 0 0 0 0 0 0 0
BG 0 1404 0 0 0 0 0 0 0 0
CG 0 0 1404 0 0 0 0 0 0 0
ABG 0 0 0 1400 0 2 2 0 0 0
BCG 0 0 0 1 1397 4 0 2 0 0
CAG 0 0 0 2 2 1398 0 0 2 0
AB 0 0 0 0 0 0 1404 0 0 0
BC 0 0 0 0 0 2 0 1402 0 0
CA 0 0 0 0 0 0 0 0 1404 0
ABC 0 0 0 0 0 0 0 0 0 1404
Table 5. Accuracy of classification.
Sl. no. Fault typeNumber of cases
Accuracy %Accuracy % Accuracy % Accuracy %
cases reported by [7] reported by [17] reported by [25]
1 LG 4212 100 100 95.3 99.72
2 LL 4212 99.95 100 99.3 100
3 LLL 1404 100 100 99.7 100
4 LLG 4212 99.59 90.61 96.8 97.50
Overall 14040 99.88 99.53* 97.77 99.17
*Based on the probability of occurrence of different types of faults.
4. Conclusions
This paper presented a new voltage transient-based fault classification technique for three terminal lines that
requires communication from the three ends for processing fault-induced voltage transients. An analysis of
transient patterns revealed the influence of FIA and Rf , and that FIA variation from 0 to 180 needs to be
considered while designing transient-based schemes. Based on investigations into fault transient patterns, a
hybrid algorithm was developed by including a PNN to resolve the misleading transient pattern problem with
LLG faults. The deterministic procedure for LG, LL, and LLL faults is very effective, with near 100% accuracy,
whereas the PNN solution for LLG faults was 99.59% accurate. The proposed method copes well with the
adverse influence of FIA and Rf . The exhaustive simulation results presented indicate the accuracy of the
scheme with all types of faults, and its speed is evident from the use of only a quarter-cycle of fault voltage
signals for processing. The method is simple, as it requires only voltage measurements, single threshold setting,
and first-level wavelet decomposition.
In practical implementation, the bandwidth of the transducers, communication requirements, and proces-
sor speed play a significant role. Optical PTs having a wide bandwidth and fiber-optic communication systems
984
BHUPATIRAJU et al./Turk J Elec Eng & Comp Sci
are needed in this scheme. Scope for further work includes hardware implementation and development of a
noise-tolerant communication-free technique.
References
[1] Carter AM, Aggarwal RK, Johns AT, Bo ZQ. Computer-aided design of a new non-unit protection scheme for EHV
teed circuits. IEE Proc-C 1996; 143: 140-150.
[2] Bhalja B, Maheswari RP. New differential protection scheme for tapped transmission line. IEE Proc-C 2008; 2:
271-279.
[3] Zhang N, Kezunovic M. Transmission line boundary protection using wavelet transform and neural network. IEEE
T Power Deliver 2007; 22: 859-869.
[4] Mahamedi M, Guo Zhu J. Fault classification and faulted phase selection based on the symmetrical components of
reactive power for single-circuit transmission lines. IEEE T Power Deliver 2013; 28: 2326-2332.
[5] Dong X, Kong W, Cui T. Fault classification and faulted-phase selection based on the initial current travelling
wave. IEEE T Power Deliver 2009; 24: 552-559.
[6] He Z, Fu L, Lin S, Bo Z. Fault detection and classification in ehv transmission line based on wavelet singular entropy.
IEEE T Power Deliver 2010; 25: 2156-2163.
[7] Valsan SP, Swarup KS. Wavelet transform based digital protection for transmission lines. Int J Elec Power 2009; 3:
379-388.
[8] Youssef ASO. Combined fuzzy-logic wavelet-based fault classification technique for power system relaying. IEEE T
Power Deliver 2004; 19: 582-589.
[9] Reddy MJ, Mohanta DK. A wavelet-fuzzy combined approach for classification and location of transmission line
faults. Int J Elec Power 2007; 29: 669-678.
[10] Vasilic S, Kezunovic M. Fuzzy ART neural network algorithm for classifying the power system faults. IEEE T Power
Deliver 2005; 20: 1306-1314.
11 Silva KM, Souza BA, Brito NSD. Fault detection and classification in transmission lines based on wavelet transform
and ANN. IEEE T Power Deliver 2006; 21: 2085-2063.
[11] Dash PK, Samantaray SR, Panda G. Fault classification and section identification of an advanced series-compensated
transmission line using support vector machine. IEEE T Power Deliver 2007; 22: 67-73.
[12] Jiang JA, Chuang CL, Wang YC, Hung CH, Wang JY, Lee CH, Hsiao YT. A hybrid framework for fault detection,
classification, and location-part i: concept, structure, and methodology. IEEE T Power Deliver 2011; 26: 1988-1998.
[13] Jamehbozorg A, Shahrtash SM. A decision-tree-based method for fault classification in single-circuit transmission
lines. IEEE T Power Deliver 2010; 25: 2190-2196.
[14] Upendar J, Gupta CP, Singh GK. Statistical decision-tree based fault classification scheme for protection of power
transmission lines. Int J Elec Power 2012; 36: 1-12.
[15] He Z, Lin S, Deng Y, Li X, Qian Q. A rough membership neural network approach for fault classification in
transmission lines. Int J Elec 2014; 61: 429-439.
[16] Livani H, Evrensoglu CY. A fault classification and localization method for three-terminal circuits using machine
learning. IEEE T Power Deliver 2013; 28: 2282-2290.
[17] Brahma SM. Distance relay with out-of-step blocking function using wavelet transform. IEEE T Power Deliver
2007; 22: 1360-1366.
[18] Chen W, Malik OP, Yin X, Chen D, Zhang Z. Study of wavelet-based ultra high speed directional transmission line
protection. IEEE T Power Deliver 2003; 18: 1134-1139.
[19] Brahma SM, De Leon PL, Kavasseri RG. Investigating the option of removing the anti-aliasing filter from digital
relays. IEEE T Power Deliver 2009; 24: 1864-1868.
985
BHUPATIRAJU et al./Turk J Elec Eng & Comp Sci
[20] Osman AH, Malik OP. Transmission line distance protection based on wavelet transform. IEEE T Power Deliver
2004; 19: 515-523.
[21] Specht DF. Probabilistic neural networks. Neural Networks 1990; 3: 109-118.
[22] Perera N, Rajapakse AD. Recognition of fault transients using a probabilistic neural network classifier. IEEE T
Power Deliver 2011; 26: 410-419.
[23] Shaik AG, Ramana Rao PV. A new wavelet based fault detection, classification and location in transmission lines.
Int J Elec Power 2015; 64: 35-40.
[24] Mollanezhad Heydar-Abadi M, Akbari Foroud A. Accurate fault classification of transmission line using wavelet
transform and probabilistic neural network. Iranian Journal of Electrical and Electronic Engineering 2013; 9: 177-
188.
[25] Coteli R. A combined protective scheme for fault classification and identification of faulty section in series compen-
sated transmission lines. Turk J Elec Eng & Comp Sci 2013; 21: 1842-1856.
[26] Guillen D, Paternin MRA, Zamora A, Ramirez JM, Idarraga G. Detection and classification of faults in transmission
lines using the maximum wavelet singular value and Euclidean norm. IET Gener Transm Dis 2015; 9: 2294-2302.
[27] Yumurtaci M, Gokmen G, Kocaman C, Ergin S, Kilic O. Classification of short-circuit faults in high-voltage energy
transmission line using energy of instantaneous active power components-based common vector approach. Turk J
Elec Eng & Comp Sci 2016; 24: 1901-1915.
986
Top Related